Thermodynamics modeling of deep learning systems for a temperature based filter pruning technique

We analyse the dynamics of convolutional filters' parameters of a convolutional neural networks during and after training, via a thermodynamic analogy which allows for a sound definition of temperature. We show that removing high temperature filters has a minor effect on the performance of the model, while removing low temperature filters influences majorly both accuracy and loss decay. This result could be exploited to implement a temperature-based pruning technique for the filters and to determine efficiently the crucial filters for an effective learning.